Drug manufacturing is the regulated path that turns an active pharmaceutical ingredient (API) and excipients into a finished medicine. The API is made and released under its own GMP guidance, ICH Q7. It is then combined with excipients and formed into a dosage form such as a tablet or a sterile vial, and finally packed and labelled. A batch reaches the market only after an authorised person has certified that it was produced and controlled according to its marketing authorisation.
What is medicine manufacturing, and what are the drug manufacturing stages?
WHO defines the finished product of medicine manufacturing as a dosage form that has been through every stage of manufacture, including packaging in its final container and labelling. The stages below follow batch manufacture; in continuous manufacture a batch is instead a defined fraction of production.
- Receipt and quarantine. Incoming containers are examined visually for correct labelling, damage, broken seals and tampering, then held in quarantine until sampled, tested and released. Q7 expects at least one identity test on each batch of material, with limited exceptions.
- API production. GMP applies from the point the company designates as the start of API production. For synthetic processes, that is the point at which API starting materials enter the process.
- API release. The quality unit reviews batch production and laboratory records for critical steps before the API batch is released or distributed. A certificate of analysis gives the batch number, the release date and each test with its result.
- Formulation. Starting materials are dispensed by designated people and independently checked. They are then blended, and milled or granulated as needed, into the final blend.
- Dosage-form making. For solid forms, the granulate is compressed into tablets or filled into capsules. For sterile products, the formulated solution is filled and sealed, then either terminally sterilised or processed aseptically throughout.
- Packaging and labelling. WHO describes packaging as the operations, including filling and labelling, that turn a bulk product into a finished product. Aseptic filling, and filling a product that will be terminally sterilised, are normally not counted as packaging.
- Finished-product quarantine and release. Finished products stay in quarantine until final release and are stored as usable stock only after that.
ICH Q10 describes a product lifecycle of pharmaceutical development, technology transfer, commercial manufacturing and product discontinuation.
How is making the API different from making the finished tablet or vial?
The API is the substance that becomes the active ingredient of the drug product and supplies its pharmacological effect. The drug product is “the dosage form in the final immediate packaging intended for marketing”. They are made under different rulebooks.
API manufacture falls under ICH Q7, which defines manufacturing as receipt of materials, production, packaging, labelling, quality control, release, storage and distribution of APIs. Q7 says GMP stringency should increase as the process moves from early API steps to final steps, purification and packaging. Physical processing of the API itself, such as milling or micronising, is also covered by Q7. The guidance stops at a defined line for sterile APIs: it applies only up to the point just before the API is rendered sterile. Sterilisation and aseptic processing follow the GMP rules for drug products.
In the US, drug products fall under the CGMP regulations in 21 CFR parts 210 and 211.
An API plant controls chemistry or biology: impurity profiles, which Q7 expects to be established for each API and confirmed during process validation, and, for fermentation or cell culture, bioburden and viral safety. A dosage-form plant controls physical form and delivery: blend uniformity, particle size, flow and compaction for solids, and sterility and particulates for injectables. EU GMP Chapter 5 links the two by requiring that supply chain traceability, from active substance starting materials to the finished medicinal product, be established, risk-assessed and periodically verified. For how AI supports the wider chain beyond the plant, see our overview of AI in supply chain and logistics.
How does manufacturing differ between tablets and sterile injectables?
Tablets and sterile injectables differ in their main process risk. For a tablet, the risk is demixing, poor flow or poor compaction. For an injectable, it is microbial, particulate or endotoxin contamination in the final product.
| Aspect | Oral solid (tablet) | Sterile injectable |
|---|---|---|
| Main process risk | Demixing, poor flow, poor compaction, non-uniform dose | Microbial, particulate and endotoxin contamination in the final product |
| Key unit operations | Blending, milling, granulation, compression | Component sterilisation, sterile filtration, filling, sealing, sometimes lyophilisation |
| Environment | Areas laid out in operation sequence at the required cleanliness levels | Critical zone at ISO 5 / Class 100 (FDA) or Grade A (EU Annex 1) for filling and closing |
| Sterility assurance | - | Terminal sterilisation where possible; otherwise aseptic processing validated by media fills |
| Unit inspection | On-line packaging checks on appearance, completeness, correct materials and over-printing | Every filled container inspected individually for contamination and defects |
| Extra batch tests | Laboratory tests against specification | Sterility and pyrogen testing for each batch |
The sterile column has two branches. In terminal sterilisation, product is filled and sealed under high-quality conditions and then sterilised in its final container by heat or irradiation. In aseptic processing, product, container and closure are sterilised separately and brought together, so there is no final sterilisation step to rescue a filling error. EU Annex 1 (2022) prefers terminal sterilisation wherever the product allows it, because it gives greater sterility assurance than sterile filtration or aseptic processing. FDA’s aseptic guidance recommends at least three consecutive successful media fill runs for initial line qualification, then semi-annual requalification for each line.
Where are the quality checks between API and a patient-ready pack?
Quality checks sit at receipt, during production and before release, and release depends on reviewing the batch records.
In-process controls are checks made during production to monitor the process, adjust it if needed and confirm the material meets its specification. Q7 expects critical in-process controls to be written down and approved by the quality unit. Production staff may adjust the process without prior approval only within limits the quality unit has already set. WHO counts control of the environment and equipment as part of in-process control too.
Batch records capture dates, major equipment, batch numbers of materials used, actual values of critical process parameters, sampling, in-process and laboratory results, and the signatures of the people who performed and checked each critical step. Every deviation, investigation and out-of-specification report is reviewed as part of batch record review before release.
The quality unit is independent of production and carries both QA and QC responsibilities. It releases or rejects APIs, raw materials, intermediates, packaging and labelling materials. For the finished product, WHO’s assessment covers production conditions, in-process results, manufacturing and packaging documentation, and compliance with the finished-product specification.
Sterile-specific checks add environmental monitoring, which FDA calls one of the most important laboratory controls in aseptic processing. They also add media fills, and inspection of the final sealed product so that damaged or defective units are removed and product that may lack container closure integrity is not shipped.
Packaging and labelling controls are designed to prevent mix-ups. Under 21 CFR 211.130, each drug product carries a lot or control number that lets the batch’s manufacturing and control history be traced. Packaging and labelling materials are examined for correctness before use. The packaging line is inspected immediately before use to confirm that product and materials from the previous run are gone, and the result is documented in the batch record. EU GMP adds yield checks and reconciliation of quantities so that discrepancies outside acceptable limits are caught.
ICH Q10 calls such a planned set of controls a control strategy. It is a planned set of controls covering materials, equipment conditions, in-process controls and finished-product specifications, with defined methods and frequencies of monitoring.
Which jobs on a drug production line does AI already handle?
One use of AI on the pharmaceutical line is automated visual inspection (AVI). AVI uses cameras, controlled lighting and software to inspect vials, syringes, tablets and packaging for cracks, discolouration, foreign particles, seal faults and incorrect labelling, and removes defective units from the line. Körber, an equipment vendor, describes AVI as combining high-resolution imaging with artificial intelligence. Applying machine learning to AVI of injectables is newer: a 2023 PDA Journal paper sets out points to consider for it, and describes the goal as capturing the judgement of human inspectors while keeping machine throughput and consistency.
TechnoLynx works on the engineering side of this: building the vision models, making them run at line speed on the inspection hardware, and producing the evidence a validation package needs. Release of the finished product stays with the manufacturer’s authorised person. Our AI engineering for life sciences and healthcare page sets out how that work is scoped.
FDA’s 2021 draft guidance on visible particulates compares automated and manual inspection. Automated inspection may improve detectability because machine variability is generally easier to control than the variability people bring to a repetitive task. In some cases it detects particulates at the low end of the visible range with greater statistical reliability. The same guidance lists high-speed cameras, diode arrays, X-ray, near-field radar and UV and near-infrared spectroscopy among the technologies in use. The camera, lighting and reject mechanics are a topic of their own. Our page on vision systems and PLC integration covers inline versus offline placement and reject design.
In our view, the next most defensible uses are the ones Q10 already asks for: monitoring process performance and product quality to keep the process in a state of control, and trending deviations into corrective action. The test we would apply is simple. Is there a defined acceptance criterion, a body of labelled historical data and a validation route? If any of the three is missing, the model has nothing to be qualified against. For uses outside pharma, the wider picture of AI in manufacturing is covered separately.
Which visual inspections still need a trained person?
People still inspect, both alongside automated systems and in place of them.
The USP Stimuli article on visual inspection notes that automated or semi-automated 100% inspection is often implemented, with human inspection reserved for acceptance quality limit (AQL) sampling and for re-examining units the machines reject. After 100% inspection, a sample of accepted units is reinspected under AQL plans, usually ANSI/ASQ Z1.4 or ISO 2859, which normally permit no critical defects. The AAPS PharmSciTech review of human inspection, drawing on Parenteral Drug Association surveys from 1996, 2003 and 2008, reports that 33–46% of participating companies used manual inspection. It expects manual inspection to remain the method of choice for small batches, infrequently made products, stability samples and budget-limited operations. Media-filled units are examined for microbial growth by personnel trained and experienced in inspecting them, with QC oversight when QC staff do not inspect them themselves.
Visual inspection is probabilistic: detection rises with particle size. In studies with spiked spherical reference particles, mostly polystyrene beads, the probability of detecting a particle smaller than 50 µm was under 5%. It rose to about 70% between 50 and 100 µm and passed 95% for most particles above 200 µm, according to the USP Stimuli article. USP <790> sets the viewing conditions: 2,000–3,750 lux, black and white backgrounds, no magnification and 5 seconds against each background. EU Annex 1 (2022) requires controlled illumination and background, qualified inspection rates, at least annual visual inspection qualification against the defect library, and frequent breaks. In a 2022 FDA-hosted presentation, Shabushnig gives two weeks as typical training time for a manual inspector.
One limit applies to both people and machines. Annex 1 states that visual inspection is not an acceptable container closure integrity test. Fused containers need integrity testing by validated methods, so a camera finding no visible defect does not establish seal integrity.
What does GMP validation require for an AI inspection system?
GMP validation of an AI inspection system requires documented proof that the system detects the known defects at least as well as qualified human inspection, plus the controls that keep that proof true over time. EU Annex 1 states that automated inspection should be validated to detect known defects and be “equal to, or better than, manual inspection methods,” challenged with representative samples. FDA’s 2021 draft guidance states that automated inspection machines should be validated to meet or surpass human inspection capabilities and can be qualified using training standards or artificial intelligence technology.
A validation package draws on the following requirements.
- A defect library. Annex 1 expects a maintained library capturing all known defect classes, with criticality set during qualification based on risk to the patient.
- A consistent human baseline. About 69% of companies in the 2008 PDA preliminary survey validated automated inspection against a manual baseline, and the AAPS PharmSciTech authors stress that the manual data must be consistent for that validation to mean anything.
- Equipment qualification. WHO GMP expects design, installation, operational and performance qualification, and singles out automated systems for particular attention.
- Routine calibration and checks. Automated inspection equipment must be calibrated, inspected or checked under a written programme, with records kept.
- Trending. Inspection results, defect types and reject levels should be recorded and trended statistically. Any critical defect found later in accepted units should trigger an investigation into the original inspection.
- Change control. Under ICH Q10, changes go through a change management system designed to give a high degree of assurance that they have no unintended consequences.
Retraining a model, or feeding it images from a new vial supplier, is a change to a validated system, and production data can move away from the qualification set without any code changing.
The manual baseline raises a judgement worth settling early. If the automated system is validated against people, the manual data must be consistent for that validation to be meaningful. The USP Stimuli authors describe visible particle analysis as still very subjective, because each manufacturer sets its own inspector training, training sets and size ranges. A team planning AVI on an injectable line should treat its defect library and inspector qualification data as validation inputs, and get them into order before the camera arrives.
FAQ
Where are pharmaceutical drugs made?
Licensed medicines are made only by holders of a manufacturing authorisation, whose sites are inspected regularly by national authorities, according to WHO GMP. Some products, such as penicillins and live biological preparations, need dedicated and self-contained facilities to prevent cross-contamination.
